Search

Given two video frames $X_{0}$ and $X_{n+1}$ , we aim to generate a series of intermediate frames $Y_{1}, Y_{2}, \ldots, Y_{n}$ , such that the resulting video consisting of frames $X_{0}, Y_{1}-Y_{n}, and X_{n+1}$ appears realistic to a human watcher. Such video generation has numerous important applications, including video compression, movie production, slow-motion filming, video surveillance, and forensic analysis. Yet, video generation is highly challenging due to the vast search space of possible frames. Previous methods, mostly based on video prediction and/or video interpolation, tend to generate poor-quality videos with severe motion blur. This paper proposes a novel, end-to-end approach to video generation using generative adversarial networks (GANs). In particular, our design involves two concatenated GANs, one capturing motions and the other generating frame details. The loss function is also carefully engineered to include adversarial loss, gradient difference (for motion learning), and normalized product correlation loss (for frame details). Experiments using three video datasets, namely, Google Robotic Push, KTH human actions, and UCF101, demonstrate that the proposed solution generates high-quality, realistic, and sharp videos, whereas all previous solutions output noisy and blurry results. [ABSTRACT FROM AUTHOR]